Disk diffusion breakpoint determination using a Bayesian nonparametric variation of the errors-in-variables model
Drug dilution (MIC) and disk diffusion (DIA) are the two most common antimicrobial susceptibility tests used by hospitals and clinics to determine an unknown pathogen's susceptibility to various antibiotics. Both tests use breakpoints to classify the pathogen as either susceptible, indeterminant, or resistant to each drug under consideration. While the determination of these drug-specific MIC classification breakpoints is straightforward, determination of comparable DIA breakpoints is not. It is this issue that motivates this research. Traditionally, the error-rate bounded (ERB) method has been used to calibrate the two tests. This procedure involves determining DIA breakpoints which minimize the observed discrepancies between results generated from both tests over a wide range of pathogen strains (or isolates). While simple and intuitive, this approach is very sample dependent and lacks precision. Model-based approaches were first proposed in 2000. These approaches model the underlying true relationship between the two tests and thus focuses on calibrating the probabilities of classification rather than the observed test results. Both a Bayesian parametric (2000) and a frequentist nonparametric (2008) procedure have been proposed. However, due to various computational difficulties and an absence of easy to use software for clinicians, neither approach has been adopted for use. In this thesis, we present a novel Bayesian nonparametric model that combines the strengths of the previous two model-based approaches. The resulting approach provides the flexibility of a nonparametric model to describe the true DIA/MIC relationship within a Bayesian framework in order to extract as much information as possible from the observed data. We demonstrate the strength of this approach via a series of simulation studies comparing it to the ERB and previous model-based approaches using breakpoint determination accuracy and model fit statistics as the comparison criteria. We conclude with applications to several real data sets and a discussion regarding software implementation and future work.
Craig, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our